RESEARCH - AI Dynamics

Instruction Mining: High-Quality Data Selection for LLMs

By

–

13 July 2023 6h33

Instruction Mining: High-Quality Instruction Data Selection for Large Language Models paper page: https://
huggingface.co/papers/2307.06
290
… Large language models typically undergo two training stages, pretraining and finetuning. Despite that large-scale pretraining endows the model with strong

→ View original post on X — @_akhaliq

13 July 2023

Robots Learning Manipulation from Human Video Demonstrations with Eye-in-Hand Cameras

By

@_akhaliq

–

13 July 2023 6h24

Giving Robots a Hand: Learning Generalizable Manipulation with Eye-in-Hand Human Demonstrations paper page: https://
huggingface.co/papers/2307.05
959
… Eye-in-hand cameras have shown promise in enabling greater sample efficiency and generalization in vision-based robotic manipulation.

→ View original post on X — @_akhaliq

13 July 2023

NaViT: Vision Transformer for Flexible Aspect Ratios and Resolutions

By

@_akhaliq

–

13 July 2023 6h20

Patch n' Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution paper page: https://
huggingface.co/papers/2307.06
304
… The ubiquitous and demonstrably suboptimal choice of resizing images to a fixed resolution before processing them with computer vision models has not yet been

→ View original post on X — @_akhaliq

13 July 2023

Sarcastic critique of AI alignment solution timeline claims

By

@rasbt

–

13 July 2023 6h18

Congrats! You should join forces with the other company who just announced that they will solve AI alignment in <4 years. So you can both work on “solving the universe” once your are done with these more near-term projects.

→ View original post on X — @rasbt

13 July 2023

Claude 2 outperforms Bard and GPT 3.5 for article summarization

By

@paulroetzer

–

13 July 2023 5h17

I’ve been experimenting with Claude 2 to summarize some of your articles and develop draft questions for your MAICON session. It’s impressive. GPT 4 performs slightly better, but Claude 2 is clearly ahead of Bard, Pi, and GPT 3.5 for this use case.

→ View original post on X — @paulroetzer

13 July 2023

Stack More Layers Differently: High-Rank Training Through Low-Rank Updates

By

@_akhaliq

–

13 July 2023 4h05

Stack More Layers Differently: High-Rank Training Through Low-Rank Updates paper page: https://
huggingface.co/papers/2307.05
695
… Despite the dominance and effectiveness of scaling, resulting in large networks with hundreds of billions of parameters, the necessity to train overparametrized models

→ View original post on X — @_akhaliq

13 July 2023

SITTA: Semantic Image-Text Alignment for Image Captioning

By

@_akhaliq

–

13 July 2023 3h57

SITTA: A Semantic Image-Text Alignment for Image Captioning paper page: https://
huggingface.co/papers/2307.05
591
… Textual and semantic comprehension of images is essential for generating proper captions. The comprehension requires detection of objects, modeling of relations between them, an

→ View original post on X — @_akhaliq

13 July 2023

Comparison Harms Collaboration in Competitive Research Environments

By

@mit_csail

–

13 July 2023 1h57

In a competitive environment where everyone is succeeding, comparison is not only going to hurt you. It will make collaboration and one of the best parts of research–asking interesting questions–painful

→ View original post on X — @mit_csail

13 July 2023

PhD Collaboration: Key to Success in Doctoral Studies

By

@mit_csail

–

13 July 2023 1h44

You can certainly collaborate with other students. That's one of the best ways to get through the PhD–by finding people that will go through the grind with you!

→ View original post on X — @mit_csail

13 July 2023